What is a reasoning model? When 'thinking first' is worth it

Two answers to the same tricky bug, below. One model answered straight away; the other worked it through first.

The second one used reasoning, often called extended thinking: it wrote out working, then answered.

Thinking is made of tokens too. Anthropic's extended thinking docs describe a budget_tokens setting that "sets a target for how many tokens Claude can use for its internal reasoning process."

More isn't free. The docs say higher budgets "enable more comprehensive reasoning, with diminishing returns that depend on the task, and at the cost of increased latency."

And you pay for it: the response reports "how many of the billed output tokens were internal reasoning." Thinking is billed as output.

The thinking also fills the context window, so long thinking in a long session uses up room faster.

Try it with the rules off. In the sandbox, give a simple rename task a huge thinking budget, then give a tricky bug no thinking at all. Compare what you paid in time for each.

Next in how the AI thinks: What are tools and agents?.